lots of efforts to accelerate training

[erow]

Key changes:

1. use crop decode strategy: see in Towards Pretraining Masked Autoencoders in One Day ! this needs turbo-jpeg 3.0
2. save memory allocation, see https://github.com/erow/ffcv/blob/main/ffcv/fields/rgb_image.py#L107
3. transformations

Comparison of throughput:

img_size	112	160	192	224					512
batch_size	512	512	512	128	256	512			512
num_workers	10	10	10	10	10	5	10	20	10
loader
ours	23024.0	19396.5	16503.6	16536.1	16338.5	12369.7	14521.4	14854.6	4260.3
ffcv	16853.2	13906.3	13598.4	12192.7	11960.2	9112.7	12539.4	12601.8	3577.8

Comparison of memory usage:

img_size	112	160	192	224					512
batch_size	512	512	512	128	256	512			512
num_workers	10	10	10	10	10	5	10	20	10
loader
ours	9.0	9.8	11.4	5.8	7.7	11.4	11.4	11.4	34.0
ffcv	13.4	14.8	17.7	7.6	11.0	17.7	17.7	17.7	56.6